Low reported severity Media Coverage

Case report: man self-treated with colchicine after a DeepSeek dialogue pointed to a rare fever syndrome; liver enzymes rose before specialist review

Clinicians at the University Medical Center Hamburg-Eppendorf report in EULAR Rheumatology Open (published 29 September 2026) a man in his thirties with nearly three decades of unexplained febrile attacks who, in a dialogue with DeepSeek begun on 28 June 2025, saw the model's differential narrow to familial Mediterranean fever (FMF). The authors write that the model repeatedly recommended specialist assessment but also discussed a diagnostic trial of colchicine with specific dosing. Motivated by the hypothesis, he arranged commercial genetic testing and obtained colchicine from a pharmacy in Egypt, starting it without specialist supervision. His attacks stopped, but his alanine aminotransferase rose to 160 U/L and colchicine was paused and later resumed at a lower dose. FMF was confirmed when he presented to the clinic in February 2026. The authors write that the case illustrates how patient-facing models may facilitate recognition of rare autoinflammatory diseases, and that confident diagnostic language combined with actionable treatment advice may also encourage unsupervised testing and self-medication.

AI system: DeepSeek; DeepSeek
Occurred: 28 Jun 2025
Reported: 29 September 2026
Event location: Egypt

What the AI did: Communicated with the person
Reported harm: Medical Deterioration

Whose AI use: Their own AI use
Setting: Health

Evidence: AI involvement reported · Causal attribution alleged · 1 source; 6 claims: 1 documented, 5 reported. 3 open questions
People reported harmed: 1 person

AI system as recorded: DeepSeek (DeepSeek-R1-0528, the publicly available model), used by the patient in an iterative diagnostic dialogue (case report by the treating clinicians)

What Happened

The case report (Department of Medicine III, University Medical Center Hamburg-Eppendorf; open access under CC BY; patient consent obtained) describes a man born in Egypt who had experienced recurrent attacks of fever, migratory abdominal pain, joint and muscle pain and skin lesions since early childhood, with C-reactive protein repeatedly above 100 mg/L during attacks, and for whom extensive gastrointestinal, urological and imaging work-up had produced no unifying diagnosis. On 28 June 2025 he began a patient-led dialogue with the publicly available DeepSeek-R1-0528 model, disclosing clinical details over successive turns; the full conversation is a supplement to the report. The authors write that the model's initial differential was dominated by urological, infectious and gastrointestinal causes but included FMF, that it later presented FMF as the leading diagnosis in increasingly confident language, and that it repeatedly discussed a diagnostic trial of colchicine and gave specific dosing suggestions before any specialist assessment, while also recommending rheumatological assessment and, in its final response about his other long-term medication, advising that colchicine be started only with specialist approval and laboratory monitoring. Motivated by the model's hypothesis, he obtained commercial MEFV genetic testing at a laboratory in Egypt, which identified two pathogenic variants, and then obtained colchicine himself from a pharmacy in Egypt and started 0.5 mg twice daily without specialist supervision. His febrile attacks ceased for the first time since childhood, but alanine aminotransferase rose to 160 U/L; colchicine was temporarily stopped while his primary care physician monitored liver enzymes, then resumed at 0.5 mg daily under close supervision. He presented to the authors' autoinflammatory disease clinic in February 2026, where FMF was confirmed on phenotype, family history, genotype and treatment response. The authors write that the case illustrates how patient-facing models may facilitate recognition of rare autoinflammatory diseases, that confident diagnostic language combined with actionable treatment advice may also encourage unsupervised testing and self-medication, and that generic advice to seek specialist advice may not sufficiently mitigate that risk.

Reported harm

The treating clinicians report that after the DeepSeek dialogue the patient started colchicine without specialist supervision and his alanine aminotransferase rose to 160 U/L, so the drug was paused and resumed at a lower dose; they also report that his attacks stopped and that the diagnosis was later confirmed.

Outcome

Unknown

Published as a peer-reviewed case report: Pauz, Kremer, Xiong and Krusche, 'LLM-supported self-diagnosis and self-treatment of familial Mediterranean fever: a case report', EULAR Rheumatology Open 2026, article 100244, doi 10.1016/j.ero.2026.100244, PMCID PMC13651492. The authors report that the patient has remained free of attacks under regular clinical and laboratory follow-up.

What remains unknown

What DeepSeek replied in each turn; the supplementary transcript was not inspected.
Where the patient held the dialogue and where his liver enzymes were monitored.
Whether the liver-enzyme rise had any cause other than colchicine, including the other long-term medication he was taking; the report does not discuss alternatives.

What the evidence supports

AI involvement: reported. The treating clinicians, who include the full conversation as a supplement, report that DeepSeek-R1-0528 progressively presented familial Mediterranean fever as the leading diagnosis in increasingly confident language and repeatedly discussed a diagnostic colchicine trial with specific dosing before specialist assessment, and that the patient, motivated by that hypothesis, arranged genetic testing and started colchicine unsupervised, after which his liver enzymes rose. They also report that the model repeatedly advised specialist supervision.

6 claims: 1 documented, 5 reported. What the statuses mean

Reported A man in his thirties had had recurrent febrile attacks since early childhood, and nearly three decades of investigations had produced no diagnosis.

Causal attribution. The patient's history as taken and written up by the treating clinicians.

ebi.ac.uk(opens in new tab) supports · English
'Since early childhood, he had experienced recurrent inflammatory attacks characterized by fever, severe migratory abdominal pain, arthralgia involving the knees, feet and hips, lower-limb myalgia, and intermittent erythematous skin lesions.'; 'Despite almost 3 decades of symptoms, extensive investigations, including contrast-enhanced abdominal CT and MRI, gastroscopy, colonoscopy, and urological assessment, failed to establish a diagnosis.'

Reported In a dialogue with DeepSeek-R1-0528 begun on 28 June 2025, the model came to present familial Mediterranean fever as the leading diagnosis in increasingly confident language and repeatedly discussed a diagnostic trial of colchicine with specific dosing before specialist assessment, while also recommending specialist evaluation and, in its final reply on his other medication, advising colchicine only with specialist approval and monitoring.

Causal attribution. The authors' description of the conversation, which they include as a supplement that was not inspected here.

ebi.ac.uk(opens in new tab) supports · English
'the patient began a dialogue with the publicly available LLM (DeepSeek-R1-0528) on June 28, 2025.'; 'FMF was presented as the leading diagnosis'; 'Although the model continued to recommend rheumatological assessment and MEFV testing, its language became increasingly confident, at times describing the presentation as classic or highly suggestive of FMF. It also repeatedly discussed a diagnostic trial of colchicine and provided specific dosing suggestions before specialist assessment.'; 'it explicitly advised that colchicine should only be initiated with specialist approval and appropriate laboratory monitoring.'

Reported Motivated by the model's hypothesis, he arranged commercial genetic testing in Egypt, which found two pathogenic MEFV variants, and obtained colchicine from a pharmacy in Egypt, starting 0.5 mg twice daily without specialist supervision.

Causal attribution. The patient's account as reported by the treating clinicians; the authors state the motivation.

ebi.ac.uk(opens in new tab) supports · English
'Motivated by the LLM-generated diagnostic hypothesis, the patient independently obtained commercial genetic MEFV testing at Alfa Labs, Egypt, which identified 2 pathogenic MEFV variants'; 'Before specialist assessment, he obtained colchicine himself from a pharmacy in Egypt and initiated treatment at a dose of 0.5 mg twice daily without prior specialist supervision.'; 'the patient nevertheless translated the generated diagnostic hypothesis into autonomous genetic testing and self-medication.'

Reported After he started colchicine his febrile attacks ceased, but his alanine aminotransferase rose to 160 U/L; colchicine was temporarily discontinued under primary-care monitoring and later resumed at 0.5 mg daily under close supervision.

Causal attribution. Clinical course reported by the authors; the liver tests were monitored by his primary care physician.

ebi.ac.uk(opens in new tab) supports · English
'Following colchicine initiation, recurrent febrile attacks ceased completely for the first time since childhood. However, alanine aminotransferase increased to 160 U/L. In view of the elevated liver enzymes, colchicine was temporarily discontinued, and liver enzyme levels were monitored closely by the primary care physician.'; 'colchicine treatment was resumed at a dose of 0.5 mg/d under close supervision.'

Documented At the authors' clinic in February 2026 the diagnosis of familial Mediterranean fever was confirmed, and he has remained free of attacks under follow-up.

Causal attribution. The treating specialists' own diagnostic finding, directly recorded in their report.

ebi.ac.uk(opens in new tab) supports · English
'Based on the characteristic lifelong clinical phenotype, Mediterranean ancestry, suggestive family history, identification of 2 pathogenic MEFV variants, and clinical response to colchicine, the diagnosis of FMF was confirmed'; 'The patient has remained free of attacks with regular clinical and laboratory follow-up.'

Reported The authors write that, although the model consistently recommended specialist evaluation, the patient turned its hypothesis into autonomous testing and self-medication, that the liver-enzyme rise shows accurate suggestions can encourage harmful behaviour when the model seems authoritative, and that generic advice to see a specialist may not mitigate risk alongside confident language and actionable treatment instructions.

Causal attribution. The authors' interpretation of the model's role; the model also advised specialist supervision, so causation is alleged.

ebi.ac.uk(opens in new tab) supports · English
'although the model consistently recommended specialist evaluation before treatment initiation, the patient nevertheless translated the generated diagnostic hypothesis into autonomous genetic testing and self-medication. The subsequent elevation of liver enzymes illustrates that even clinically accurate diagnostic suggestions may inadvertently encourage harmful behavior if patients perceive the model as sufficiently authoritative.'; 'generic recommendations to seek specialist advice may not sufficiently mitigate risk when they are accompanied by confident diagnostic language and actionable treatment instructions'

Sources

1 source inspected. Sources that repeat one account do not corroborate each other.

EULAR Rheumatology Open, 29 September 2026: Pauz et al., LLM-supported self-diagnosis and self-treatment of FMF (case report, Europe PMC full text)(opens in new tab)
s1 · ebi.ac.uk · Clinical case report · English · Inspected · 29 September 2026 · Primary

How the sources were read, and where the events happened

Full text of the peer-reviewed case report read as Europe PMC fullTextXML for PMC13651492 (HTTP 200) on 2026-10-12: abstract, case description, discussion, consent and AI-use declarations. Written by the treating clinicians. The supplementary DeepSeek conversation was not inspected. Applies to s1.

Event countries: Egypt. Affected-person countries: Unknown. Court countries: Unknown.

The report says the patient obtained the genetic test at a laboratory in Egypt and the colchicine from a pharmacy in Egypt (EG). It does not say where he held the DeepSeek dialogue or where his liver enzymes were monitored. He later presented to the authors' clinic in Hamburg; that confirmation visit is not used as the event location, and birthplace is not used as residence. No court proceeding is reported.

Reviewed for publication 2026-10-12: Core case (communicated_with): the treating clinicians report that a DeepSeek dialogue presented a rare fever syndrome as the leading diagnosis with specific colchicine dosing, and that the patient then self-medicated without supervision and developed a liver-enzyme rise that led to a pause in treatment. The account is mixed: the model also advised specialist supervision and the patient's attacks stopped, so causation is alleged and severity low.

People described

A man in his thirties with long-standing recurrent fevers, described in a case report in a rheumatology journal published with his written consent; not named in the report or here.

People reported harmed in this case

1 person

1 AI participant · 0 other people harmed

One patient, the subject of the case report. Exact 1.

Counted once within this case. The same person may appear in other cases. This count does not establish AI causation.

Cite this case

Compiled per our published methodology: verification statuses, sourcing standards, and corrections process.

APA

NOPE. (2026). Case report: man self-treated with colchicine after a DeepSeek dialogue pointed to a rare fever syndrome; liver enzymes rose before specialist review. AI incidents. https://nope.net/incidents/2025-case-report-deepseek-guided-fever-self-diagnosis-unsupervised-colchicine-liver-enzymes

BibTeX

@misc{2025_case_report_deepseek_guided_fever_self_diagnosis_unsupervised_colchicine_liver_enzymes,
  title = {Case report: man self-treated with colchicine after a DeepSeek dialogue pointed to a rare fever syndrome; liver enzymes rose before specialist review},
  author = {NOPE},
  year = {2026},
  howpublished = {AI incidents},
  url = {https://nope.net/incidents/2025-case-report-deepseek-guided-fever-self-diagnosis-unsupervised-colchicine-liver-enzymes}
}

Wuhan, China: reports say a family ate wild mushrooms an AI identification tool called non-poisonous and a 9-year-old fell critically ill; one cites amatoxin

A family from Wuhan, Hubei province, gathered wild mushrooms resembling the edible termite mushroom (jizong) during a mountain outing in June 2026 and, according to a Guangming Online post carried by The Paper on 5 July 2026 and an Asahi Shimbun report of 6 July citing the Changjiang Daily, ran an AI identification tool on them. The tool returned a "non-poisonous" result and the family cooked and ate the mushrooms. According to the Guangming Online post, all three family members vomited soon after the meal and again before dawn the next day; they were admitted to Zhongnan Hospital of Wuhan University, where the 9-year-old child was moved to the paediatric intensive care unit the same afternoon. The post says the child had amatoxin poisoning that progressed to the most severe grade, with acute liver failure, hepatic encephalopathy and fluid in the chest and abdomen, and went home after eight days in hospital. The AI tool is not named, its output is not reproduced, and the account comes from one reporting chain.

High Unidentified mushroom-identification tool

Zhuji, Zhejiang: reports say a man ate wild mushrooms after an AI photo check called them non-toxic; rhabdomyolysis and acute kidney failure treated in an ICU

A man in Zhuji, Zhejiang province, gathered wild mushrooms near his home, photographed them and asked an AI tool whether they were poisonous, according to a Science and Technology Daily item carried by Qingdao Daily on 20 August 2026 (citing Zhuji Daily) and reports by the World Journal (20 August) and Sing Tao (21 August) citing Dushi Kuaibao's Chengshi app. The reports say the tool answered that the mushrooms were non-toxic and he cooked and ate them at noon. That afternoon he developed severe abdominal pain, repeated watery diarrhoea and aching arms; at Zhuji People's Hospital his creatine kinase was reported at 28,771 U/L, about 100 times the normal value, with acidosis and raised lactate, and doctors diagnosed rhabdomyolysis and acute kidney failure. He was moved to the intensive care unit, treated to remove the toxin and protect organ function, and has since been discharged. The AI tool is not named, its output is not reproduced, and the account rests on one hospital-and-local-press account relayed by two chains.

Critical Perplexity

Son reports his father delayed cancer treatment while relying on Perplexity

A man with cancer delayed recommended treatment while relying on a Perplexity report, according to his son and subsequent New York Times reporting. The newspaper interviewed his oncologist and a researcher whose work the report misrepresented. He died in December 2025. His son expressly cautioned that AI did not simply “kill” his father and that he might have resisted treatment without it.

Medium Unidentified mushroom-identification tool

Wakayama, Japan: city health officials say a man in his 70s was hospitalised after an AI photo check told him a poisonous wild mushroom was edible (MBS report)

According to Wakayama city’s public-health division, as reported by MBS (Mainichi Broadcasting) on 21 November 2025 and relayed by HuffPost Japan on 9 September 2026, a man in his 70s living in Wakayama city gathered wild mushrooms resembling hiratake (oyster mushroom) and shiitake in the mountains of Shimokitayama village, Nara prefecture, on 3 November 2025. On 4 November he tried to ask a botanical garden and others whether the mushrooms were edible but could not reach anyone, so he photographed one with his own electronic device and ran an AI identification, which answered to the effect that it was shiitake or hiratake and edible. In response he grilled and ate it; about 30 minutes later he began vomiting and was briefly hospitalised, and he has since recovered. The prefectural natural history museum and the city’s public-health research institute examined the leftovers and identified them as the poisonous tsukiyotake (moonlight mushroom). The city warned that judging wild mushrooms by AI or by picture books alone is dangerous. HuffPost Japan reports that Japan’s agriculture ministry issued a warning on 4 September 2026 against deciding edibility from search results or AI answers. The AI tool is not identified in the reports and the man’s exact exchange with it is not reproduced.

If you or someone you know is struggling, free and confidential support is available. Find a helpline near you at Signpost.

Back to all incidents